Diffeomorphic Sulcal Shape Analysis for Cortical Surface ... · proaches. Their shapes have largely...
Transcript of Diffeomorphic Sulcal Shape Analysis for Cortical Surface ... · proaches. Their shapes have largely...
Diffeomorphic Sulcal Shape Analysis for Cortical Surface Registration
Shantanu H. Joshi
Ryan P. Cabeen
Anand A. Joshi
Roger P. Woods
Katherine L. Narr
Arthur W. Toga
Laboratory of Neuro Imaging, University of California Los Angeles
Los Angeles, California, 90095, USA
Abstract
We present an intrinsic framework for constructing sul-
cal shape atlases on the human cortex. We propose the anal-
ysis of sulcal and gyral patterns by representing them by
continuous open curves in R3. The space of such curves,
also termed as the shape manifold is equipped with a Rie-
mannian L2 metric on the tangent space, and shows desir-
able properties while matching shapes of sulci. On account
of the spherical nature of the shape space, geodesics be-
tween shapes can be computed analytically. Additionally,
we also present an optimization approach that computes
geodesics in the quotient space of shapes modulo rigid rota-
tions and reparameterizations. We also integrate the elas-
tic shape model into a surface registration framework for
a population of 176 subjects, and show a considerable im-
provement in the constructed surface atlases.
1. Introduction
The cerebral cortex in mammalian brains is largely re-
sponsible for memory, thought processes, and sensory per-
ceptions, and plays a key role in social abilities, language
development, and abstract reasoning and thought, espe-
cially in human brains. The cortex comprises of neurons
and nerve fibers and appears gray in color. The boundary of
this gray matter region, also referred to as the cortical sur-
face, can be observed as a highly convoluted exterior with
valleys (sulci) and ridges (gyri). Cortical surfaces among
populations exhibit highly varied folding patterns and it is
of great interest to neuroscientists as well as neurologists
to recognize, classify, and understand them. A surface-
based morphometric analysis of the cortex is shown to have
wide reaching applicability for the purpose of mental dis-
ease detection, progression, as well as prediction and un-
derstanding of normal and abnormal developmental behav-
iors. Surface-based morphometry has three major ingredi-
ents: i) representation of the cortical surface, ii) registration
and alignment of one or more surfaces for construction of
atlases, and iii) statistical analysis of deformations or warps
explaining the variability of surface features in a given pop-
ulation. Surface registration aims at determining correspon-
dences between a pair of surfaces by aligning several ho-
mologous features with respect to each other. The corre-
spondence can be achieved either automatically, for e.g. us-
ing both local and global features as in the case of Fischl et
al. [5, 2], or in a semi-automated manner, using expertly de-
lineated sulcal and gyral landmarks as in the case of Thomp-
son et al. [14, 7]. These landmarks are usually 3D contin-
uous space curves corresponding to the deepest regions of
the valleys for sulci, and topmost regions of the ridges for
the gyri. Typically it is difficult for automated algorithms
to locate, and classify sulcal and gyral patterns exclusively
based on local geometric features. Thus the main advan-
tage of using explicit landmarks is the incorporation of ex-
pert anatomical knowledge that improves the consistency in
matching of homologous features. This in turn potentially
improves statistical power in the neighborhood of the land-
marks. Additionally, increasing the number of consistent
landmarks also improves the alignment accuracy, thereby
allowing more control in the registration process.
Previously, landmark curves have mostly been used
as boundary conditions for various cortical alignment ap-
proaches. Their shapes have largely been ignored during
registration. The sulcal and gyral fissures are highly cor-
related with the folding patterns of the cortex. Recently,
there has been a renewed interest in studying the cortical
folding patterns for the purpose of increased understand-
ing of the organizational processes leading to the develop-
ment of the cortex [4, 10, 3, 12]. Furthermore, the shapes of
such patterns are also linked with brain function. In this pa-
per we focus on modeling the shape of sulci and gyri, and
978-1-4244-6983-3/10/$26.00 ©2010 IEEE
incorporating it during the process of cortical registration
and atlas construction. In the past, various researchers have
modeled the sulci and gyri using different representations.
Leow et al. [9] have demonstrated multiple curve compar-
ison strategies that include homothetic mapping of curves
discretized with equally spaced points, and whole-curve
matching by relaxing point-wise correspondences between
pairs of curves. However, they project three-dimensional
sulcal geometry onto a two-dimensional plane, and then
use diffeomorphic mappings between the flat planes to find
point correspondences. Tao et al. [13] represent sulci using
landmark points on curves, and build a statistical model by
using a Procrustes alignment of sulcal shapes. Vaillant et
al. [15] represent cortical sulci by medial surfaces of corti-
cal folds. They subsequently parameterize the surface into
salient isoparametric active contours. Although, the advan-
tage of this model is that it represents entire cortical folds, a
limitation of this method is the use of unit speed parameteri-
zations of active contours for constructing Procrustes shape
averages for sulci. This implies that the average shape is
only invariant to translation, rotation, and scaling. Further-
more, for the latter two approaches, the shapes are repre-
sented by finite features or landmarks, and thus are limited
in characterization of the rich geometric detail that mani-
fests in the cortical folds that give the sulci their shapes.
1.1. Contributions
In this paper, unlike the previous approaches, we
will represent the sulci and gyri by parameterized three-
dimensional curves. Instead of analyzing the curves di-
rectly, we will deal with their functional representations that
map curve instances onto a shape manifold. The main con-
tributions of this paper are as follows:
1. An inverse-consistent diffeomorphic framework for
matching sulcal shapes.
2. An intrinsic sulcal shape atlas based on the Rieman-
nian metric on the shape manifold.
3. Integration of sulcal curve diffeomorphisms in driving
cortical surface registrations.
To our knowledge, the proposed framework of direct diffeo-
morphic three-dimensional sulcal curve mappings have not
been used in cortical registration before. Additionally, we
also show comparisons of shape averages of our approach
to the standard Euclidean curve matching that is normally
used in surface registration.
The remainder of this paper is organized as follows. Sec-
tion 2 outlines the shape modeling scheme for sulci and
gyri. It deals with the shape representation, and specifies a
Riemannian metric on the tangent space of the shape man-
ifold. Section 3 outlines the procedure for computing sta-
tistical shape averages for sulci and gyri for a given popula-
tion. Finally Section 4 incorporates the sulcal shape model
in cortical surface registration, followed by results and con-
clusion.
2. Diffeomorphic Shape Analysis of Sulci and
Gyri
In this section, we describe the modeling scheme used to
represent sulcal and gyral shape features. We will represent
the cortical valleys (sulci), and the ridges (gyri) by open
curves. However unlike previous approaches which have
used landmarks for representing the sulcal and gyral fea-
tures, we will use continuous functions of curves for repre-
senting shapes. For example, Fig. 1 shows the lateral, me-
dial and the ventral views of a cortical surface along with
manually highlighted sulci and gyri.
Figure 1. Cortical Surface with manually traced sulci and gyri.
2.1. Shape Representation
We use a vector valued map to parameterize the 3D sulci
and gyri as follows. Let β be a 3D, arbitrarily parameter-
ized, open curve, such that β : [0, 2π] → R3. We represent
the shape of the curve β by the function q : [0, 2π] → R3 as
follows,
q(s) =β(s)√||β(s)||
∈ R3 . (1)
Here, s ∈ [0, 2π], || · || ≡√
(·, ·)R3 , and (·, ·)R3 is taken to
be the standard Euclidean inner product in R3. The quantity
||q(s)|| is the square-root of the instantaneous speed, and the
ratioq(s)
||q(s)|| is the instantaneous direction along the curve.
The original curve β can be recovered upto a translation,
using β(s) =∫ s
0 ||q(t)|| q(t) dt. In order to make the rep-
resentation scale invariant, we will normalize the function qas q = q√
(q,q)R3
. We now denote Sq ≡ q|q(s) : [0, 2π] →
R3|
∫ 2π
0 (q(s), q(s))R3ds = 1 as the space of all unit-
length, elastic curves. On account of scale invariance, the
space Sq becomes an infinite-dimensional unit-sphere and
represents all open elastic curves invariant to translation and
uniform scaling. The tangent space of Sq is easy to define
and is given as Tq(Sq) = w = (w1, w2, . . . , wn)|w(s) :
I → R3 ∀s ∈ [0, 2π) |
∫ 2π
0 (w(s), q(s))R3 ds = 0. Here
each wi represents a tangent vector in the tangent space of
Sq. We define a metric on the tangent space as follows.
Given a curve q ∈ Sq, and the first order perturbations of
q given by u, v ∈ Tq(Sq), respectively, the inner product
between the tangent vectors u, v to Sq at q is defined as,
〈u, v〉 =
∫ 2π
0
(u(s), v(s))R3ds. (2)
Now given two shapes q1 and q2, the translation and scale
invariant shape distance between them is simply found by
measuring the length of the geodesic connecting them on
the sphere. We know that geodesics on a sphere are great
circles and can be specified analytically. The geodesic on
Sq between the two points q1, q2 ∈ S
q along a unit direction
f ∈ Tq1(Sq) towards q2 for a time t is given by
χt(q1; f) = cos(t cos−1〈q1, q2〉
)q1+
sin(t cos−1〈q1, q2〉
)f, (3)
where t is a subscript for time. Then the distance between
the two shapes q1 and q2 is given by
d(q1, q2) =
∫ 1
0
√χt, χtdt (4)
The quantify χt is also referred to as the velocity vector
field on the geodesic path χt So far, we have constructed
geodesics between a pair of curves directly on the sphere
(Sq). In doing so, we implicitly assumed that the curves
were rotationally aligned, as well as the parameterization of
the curves was fixed. However the shape of a curve remains
unchanged under rotations as well as different parameteri-
zations of the curve. Thus in order to register shapes ac-
curately, the matching should be invariant to rotations as
well as reparameterizations. In the following section, we
define the space of elastic shapes, and outline an optimiza-
tion algorithm that measures the “elastic” distance between
curves under certain well-defined shape-preserving trans-
formations.
2.2. Geodesics between Elastic Shapes
In order to match curves elastically, in addition to trans-
lation and scaling, we consider the following reparameter-
izations and group actions on the curve that preserve its
shape. A rigid rotation of a curve is considered as a group
action by a 3 × 3 rotation matrix O3 ∈ SO(3) on q, and
is defined as O3 · q(s) = O3q(s), ∀s ∈ [0, 2π]. A curve
traveled at arbitrary speeds is said to be reparameterized
by a non-linear differentiable map γ (with a differentiable
inverse) also referred to as a diffeomorphism. We define
D = γ : S1 → S
1 as the space of all orientation-
preserving diffeomorphisms. Then the resulting variable
speed parameterizations of the curve can be thought of as
diffeomorphic group actions of γ ∈ D on the curve q. This
group action is derived as follows. Let q be the representa-
tion of a curve β. Let α = β(γ) be a reparameterization of
β by γ. Then the respective velocity vectors can be written
as α = γβ(γ) = γq(γ)||q(γ)|| = ||√γq(γ)||√γq(γ). The
reparameterization of q by γ is defined as a right action of
the group D on the set C and written as q · γ =√
γ (q γ).Thus we are interested in constructing the shape space as a
quotient space of Sq , modulo shape preserving transforma-
tions such as rigid rotations and reparameterizations.
Altogether, the set of curves affected by the group ac-
tions above, partition the space Sq into equivalent classes.
We now define the elastic shape space as the quotient space
S = Sq/(SO(3) × D). The problem of finding a geodesic
between two shapes in S is same as finding the shortest path
between the equivalent classes of the given pair of shapes.
Since the actions of the re-parametrization groups on C con-
stitute actions by isometries, this problem also amounts to
minimizing the length of the geodesic path, such that
de(q1, q2) = minO3∈SO(3),γ∈D
d(q1, (O3q2) · γ), (5)
where d is given by the distance in Eq. 4. In order to op-
timize Eq.5, we recognize that for a fixed rotation O3, the
distance de can be obtained by finding the optimal reparam-
eterization γ between q1 and q2, whereas for a fixed γ, the
distance de is calculated by finding the optimal rotation O3.
Thus in order to minimize the distance in Eq. 5, we alter-
nate between optimizing over O3 and γ repeatedly until the
process converges. At each step, the optimal rotation O3
is given by O3 = USV T , where all U, S, V ∈ R3×3, and
USV T =∫ 2π
0q1(s)q2(s)
T ds and is obtained using singu-
lar value decomposition. Also, at each iteration, we com-
pute a geodesic path given by Eq. 3 between the starting
shape q1 and the target shape O3q2 · γ. In order to find the
optimal γ, we minimize the incremental L2 cost function
∫ t
s||q1(τ)−
√˙γ(τ)q2(γ(τ))||2dτ using dynamic program-
ming. Figure 2 shows examples of curve correspondences
obtained using both non-elastic and elastic geodesics, along
with the optimal diffeomorphism γ. The trivial diffeomor-
phism γ = s in the non-elastic case is also shown for com-
parison.
3. Intrinsic Statistical Model for Sulci
Having defined the framework for curve shape match-
ing, we now construct a sulcal atlas of a large population of
Figure 3. Lateral, frontal, and medial views of, Top row: 28 landmark sulci and gyri for 176 subjects, Middle row: Euclidean sulcal shape
averages for each landmark, Bottom row: Karcher means for each landmark. (Best viewed in color).
−0.3−0.2−0.1 0 0.100.10.20.30.4
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
−0.3−0.2−0.1 0 0.100.10.20.30.4
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0 1 2 3 4 5 60
1
2
3
4
5
6
Optimum Diffeo
Identity Diffeo
−0.1
0
0.1
−0.7−0.6−0.5−0.4−0.3−0.2−0.100.1
0
0.1
0.2
0.3
0.4
0.5
−0.1
0
0.1
−0.7−0.6−0.5−0.4−0.3−0.2−0.100.1
0
0.1
0.2
0.3
0.4
0.5
0 1 2 3 4 5 60
1
2
3
4
5
6
Optimum Diffeo
Identity Diffeo
(a) (b) (c)Figure 2. Each row shows an example of: (a) Curve registration
obtained by a non-elastic geodesic using Eq. 4, (b) Curve registra-
tion obtained by an elastic geodesic using Eq. 5, and (c) optimal
diffeomorphism γ shown in blue, along with γ = s plotted in red.
curves in the shape space. However, in order to do so, we
need the notion of a shape average based on the sulcal and
gyral curves. Owing to the nonlinearity of the shape space,
the computation of an average shape is not straightforward.
There are two well known approaches of computing statis-
tical averages in such spaces. The extrinsic shape average
is computed by projecting the elements of the shape space
in the ambient linear space, where an Euclidean average
is computed, and subsequently projected back to the shape
space. On the other hand, the intrinsic average, also known
as the Karcher mean ([1, 8]) is computed directly on the
shape space, and makes use of distances and lengths that are
defined strictly on the manifold. It uses the geodesics de-
fined via the exponential map, and iteratively minimizes the
average geodesic variance of the collection of shapes. The
extrinsic average is simple and efficient to calculate, how-
ever it ignores the nonlinearity of the shape space, as well as
depends upon the method used for embedding the manifold
in the Euclidean space. As a simple example, a unit circle
can be embedded inside R2 in several ways, and each em-
bedding possibly leads to different values of extrinsic means
on the circle. We will adopt the intrinsic approach by com-
puting the Karcher mean for a given set of shapes. Algo-
rithm 1 describes the procedure for computing the Karcher
mean of a collection of shapes under the q representation.
Using Algorithm 1, we now construct a sulcal shape at-
las for a population of 176 subjects. These subjects under-
went high-resolution T1-weighted MRI scans with spatial
resolution 1x1x1 mm3 on a Siemens 1.5T scanner. Af-
ter preprocessing the raw data, and registering it stereotaxi-
cally to a standard atlas space, the cortical surfaces for these
subjects were extracted using an automated algorithm [6].
Algorithm 1: Given a collection of shapes qi, i =1, . . . , N , find their Karcher mean shape.
Initialize the mean shape as µ = 1N
∑N
i=1 qi1
Project the mean shape on the shape space Sq as2
µ = µ√(µ,µ)
R3
while not converged do3
for i = 1, . . . , N do4
Find an elastic geodesic path χ after following5
the procedure given in Sec. 2.2
Let χ(i) = dχdt
(0).6
end7
¯χ = 1N
∑Ni=1 χ(i)8
µ = χ1(µ, ¯χ).9
end10
For each of these subjects, a total of 28 landmark curves
were manually traced. Figure 3 shows the original 28 land-
mark curves for each of the 176 subjects for the left hemi-
sphere overlaid together. For the remainder of this paper,
we will demonstrate the results on the left hemisphere of
the brain to simplify things, although the same procedure
can be repeated for the right hemisphere as well. Figure
3 also shows the intrinsic sulcal shape averages of the 28
landmark curves using Algorithm 1, as well as the respec-
tive extrinsic Euclidean averages for the same. While com-
puting the extrinsic average, each curve for the same land-
mark type was mapped to its q representation, thus making
it scale and translation invariant. The Euclidean average of
all the q functions was then computed after a pairwise rota-
tional alignment. Both the Karcher mean shapes as well as
the Euclidean averages were then mapped back to the curve
space in order to visualize them.
It is observed that the intrinsic averages although
smooth, have preserved important features along the land-
marks, thus representing the average local shape geometry
along the sulci and gyri. This also implies that the shape
average has not only captured the salient geometric fea-
tures, but has also reduced the shape variability in the pop-
ulation. In order to demonstrate this property, we plot the
variance of the shape deformation for each landmark type
as captured by the velocity vector along the geodesic path,
both for Euclidean extrinsic, and Riemannian intrinsic av-
erages. Thus for each of the 28 landmark average shapes
for 176 subjects, µi, i = 1, . . . , 28, we plot the quantity1
176
∑11 76〈 ˙χ(i), ˙χ(i)〉, where ˙χ(i) given in Algorithm 1,
line 6. This quantity measures the invariant deformation
between a pair of shapes, and only depends upon the intrin-
sic geometry of the shapes. Figure 4 shows a comparison of
the plots of ˙χ(i) for each of the landmarks, taken along the
length of the curve, for both Euclidean shape averages, as
well as intrinsic shape averages. This can also be thought
of as the geodesic variance. From the color-coded map, it
is observed that the intrinsic average has reduced the vari-
ance in terms of shape geometry deformation, and thus is a
better representative of the population. This has important
implications for surface registration as is demonstrated in
the following section.
4. Cortical Surface Registration using Land-
mark Curves
In this section, we briefly introduce our scheme for reg-
istering cortical surfaces. There have been prominent ap-
proaches [14, 7] for cortical registration in the neuroimag-
ing community. The method, although based on the same
conceptual framework of elastic registration, provides a
slightly different model for computing the deformation, and
there are improvements in the implementation that increase
flexibility and efficiency.
A general outline of the process is as follows. The first
stage is to establish homology of the curve data, which is ac-
complished by matching the shapes in a Riemannian shape
space framework. Next, the surfaces and curves are confor-
mally mapped to the sphere, establishing a common space
where deformation will be defined. Following this, there
is a rotational alignment of the surfaces and curves to ac-
count for the differences in spherical mapping orientations.
Next, the spherical mean of the curves is found to define
the atlas curves for the domain of the deformation. Finally,
the elastic deformation of the atlas on the sphere is found,
which constrained to map the atlas curves to each case’s set
of curves. The surfaces are reparameterized by this elastic
deformation. The resulting surfaces then have homologous
coordinate systems, allowing local comparisons across the
group.
We define a set of N surfaces, M1, ..., MN where
Mi ⊂ R3. We represent their mesh geometry using a set of
simplicial complexes, (K1, f1), ..., (KN , fN) where Ki
is a simplicial complex and fi : K 7→ Mi. For each surface
i, we have a set of M landmarks represented by continuous
open curves βi1, ..., βiM where βij : [0, 2π] 7→ Mi, and
the set of curves βij : i ∈ [1, N ] represent homologous
regions on the set of surfaces. Additionally, the curves are
discretized, where the j-th curve has kj vertices.
4.1. Inverse Diffeomorphic Mapping of LandmarkCurves
The first step of the process is to establish a correspon-
dence between the homologous landmark curves by com-
puting mappings γij : [0, 2π] 7→ [0, 2π] such that for curve
j and parameter t, βij(γij(t)) : i ∈ [1, N ] is a set of ho-
mologous points on the surfaces. This is accomplished by
Figure 4. Comparison of the geodesic variance for the entire sulcal population for each of the 28 landmarks, both for Euclidean shape
averages, as well as elastic shape averages, along the length of the curves.
mapping the curves to a Riemannian manifold shape space,
where reparameterizations are defined by geodesics to the
Karcher mean of the curves in the shape space.
4.2. Spherical Mapping and Alignment
Next, the surfaces are mapped to the sphere, where the
deformation will be computed. In general, a set of bijec-
tive mappings is found from each surface to the unit sphere,
φ1, ..., φN where φi : Mi 7→ S2. The spherical map-
ping of the matched curves is then βij = φi βij γij .
The meshes are simplified using a QEM-based simplifi-
cation and mapped to a sphere using a conformal map-
ping. The curves are then mapped to the sphere by us-
ing the barycentric coordinates of each curve vertex. A
bounded interval hierarchy is used to efficiently search for
the coincident face of each curve vertex. Once the meshes
and curves are mapped to the sphere, they are rotationally
aligned to enforce a consistent orientation of the spherical
mappings. Given an arbitrarily chosen target, each set of
curves is aligned to the target by computing the rotation
and reflection that minimizes the least-squared difference
between the discretized curve coordinates. This is accom-
plished by solving the unconstrained orthogonal Procrustes
problem using singular value decomposition. Typically, left
and right hemisphere surfaces will be available, and allow-
ing reflections in the transform is advantageous in this case
because both hemisphere can be mapped to a common ori-
entation. So, for an arbitrary T ∈ βij : i ∈ [1, N ], we
find an alignment R ∈ R3×3 as,
Ri = arg minΩ
∑
j
∑
k
‖Ω(
βij
(k − 1
kj − 1
))− T
(k − 1
kj − 1
)‖,
(6)
where ΩT Ω = I
We then optimally rotate the curves as well as the spherical
maps as, φi = Ri φi, and βij = Ri βij .
4.3. Spherical Curve Atlas
Once the meshes and curves have been aligned on the
sphere, the mean curves are computed. These mean curves
will be used as the atlas curves in the surface warping. The
Karcher mean on the sphere is found for each vertex of each
curve, across the group. In this method, an initial guess
is found by the normalized average of the points. For this
point, the tangent space is defined by the gnomonic projec-
tion. A new mean is computed in the tangent space, and
it is mapped back to the sphere and the process is repeated
until the difference between the previous and the new mean
is smaller than a threshold. We can express the curve atlas
as the set βj : j ∈ [1, M ], where βj is the karcher mean
of βij : i ∈ [1, N ].
4.4. Elastic Surface Warping
For surface i, the deformation of the atlas is φi : S2 7→S2, where φi(βj(t)) = βij(t) for t ∈ [0, 2π], j ∈ [1, M ].Six bijective flattenings of the sphere are defined ϕn :S2 7→ [0, 1]2 : n ∈ 1, 2, 3, 4, 5, 6. For a point on
the sphere, p ∈ S2, the flattening is chosen as ϕp =argminϕn
‖ϕn(p) −(
12 , 1
2
)‖. The displacement field up :
S2 7→ R2 is then up(x) = ϕp(φi(x)) − ϕp(x). At non-
landmark points, i.e. x ∈ S2, x /∈ ∪jβij(t) : t ∈ [0, 2π],
the mapping is constrained to satisfy a small deformation
nonlinear elastic model similar to Thompson et al. [14],
and is given by
µ2up(x) + (λ + µ)( · up(x)) = 0 (7)
Figure 5. Mappings between the grid, octahedron, sphere and a
cortical surface
Computationally, the atlas mesh is defined on the sphere
by tessellating the sphere with a subdivided octahedron,
shown in Figure 5 [11]. This representation is advanta-
geous for its multiscale and flattened representation. Sub-
division of each triangle is performed by adding vertices
at the midpoints of the edges, then adding new edges con-
necting them, creating four triangles from the original. This
scheme is used for defining the representations of the mesh
in multiscale algorithms. Furthermore, this subdivision pro-
cess ensures that the mesh can be flattened to a rectangular
and regularly sampled grid. This flattening can be imagined
as follows. First, choose one of the vertices of the octahe-
dron, this will be mapped to the center of the grid. Then
cut the four far edges that do not contain the center vertex.
These edges are duplicated and define the boundary of the
grid, and the opposite vertex maps to the four corners of the
grid. This is illustrated in Figure 6. This flattened repre-
sentation allows for efficient interpolation, smoothing and
finite differences operations on the grid [11].
x
x x
xa
c
d b
x
d b
c
a x
x x
xa
c
d b
Figure 6. The flattening and an example of two subdivisions.
The deformation is computed iteratively using finite dif-
ferences with a multigrid method. The octahedral domain
is particularly suitable for this because the representation
lends to simple prolongation, restriction and smoothing
multigrid operations. The solver accounts for the spheri-
cal topology of the domain by solving the equation in the
flattening of the mesh at the vertex that is being updated,
mapping neighboring vertex values relative to this flatten-
ing, as described above. Each mesh is resampled by the
deformed atlas, establishing vertex homology between the
meshes.
5. Results
In this section, we demonstrate results of cortical sur-
face registration with and without the incorporation of the
elastic sulcal shape analysis. The data consists of cortical
surfaces of 176 subjects collected and processed according
to the procedure described in Section 3. The sulcal shape
atlas is already constructed and shown in Fig. 3. We now
compute geodesics between the average shape of the land-
mark, and the set of all sulci belonging to that landmark
type, and reparameterize the set of sulci according to in-
verses of the resulting diffeomorphisms. We then follow
the steps outlined in Sec. 4 in order to warp all the sur-
faces meshes to the atlas. Figure 7 shows the flattened rep-
resentation of a surface atlas colored according to curvature
of the surface. The figure shows flattened atlases for sur-
faces without any landmark curves, surfaces with landmark
curves using Euclidean analysis, and finally surfaces with
landmark curves using the diffeomorphic shape analysis. In
the absence of landmark curves, the atlas shows smooth-
ing of corresponding features and information present in
the population, while in the presence of landmarks, the gen-
eral landmark features reappear in the atlas. However the
most interesting case is shown in Fig. 7 (c), where not only
do the landmark curves manifest themselves as before, but
they also exhibit rich local geometry that has been preserved
due to the elastic diffeomorphisms. Additionally, Fig. 8
shows three the lateral, dorsal, and medial views of the re-
constructed cortical surface from the flattened representa-
tions. It is observed that the surface with diffeomorphic
shape analysis shows richer geometric detail than the typ-
ical Euclidean analysis.
(a) (b)Figure 8. Lateral, medial, and ventral views of the reconstructed
cortical surface with (a) Euclidean landmark analysis, and (b) dif-
feomorphic shape analysis.
(a) (b) (c)Figure 7. Average flattened representations for atlases: (a) without landmark curves, (b) with landmark curves using Euclidean analysis,
(c) with landmark curves using the diffeomorphic shape analysis.
6. Conclusion
We have presented a geometric method for shape analy-
sis of sulcal and gyral features and demonstrated their use in
cortical surface registration. Although, the method models
shape averages of cortical sulci and gyri, the theory itself is
quite general and is applicable to diverse surface correspon-
dence problems that use landmarks to guide the warping.
Thus surface registration accuracy is improved, not only by
the inclusion of increasing number of landmarks, but largely
by the proposed intrinsic, elastic shape analysis framework.
Acknowledgments
This research was partially supported by the National In-
stitute of Health through the National Center for Research
Resources (NCRR) for Center for Computational Biology
(CCB), Grant U54 RR021813.
References
[1] R. Bhattacharya and V. Patrangenaru. Nonparametic esti-
mation of location and dispersion on Riemannian manifolds.
Journal of Statistical Planning and Inference, 108(1-2):23–
35, Jan 2002.
[2] A. M. Dale, B. Fischl, and M. I. Sereno. Cortical surface-
based analysis. I. segmentation and surface reconstruction.
NeuroImage, 9(2):179–94, Feb 1999.
[3] E. Duchesnay, A. Cachia, A. Roche, D. Riviere, Y. Coin-
tepas, D. Papadopoulos-Orfanos, M. Zilbovicius, J.-L. Mar-
tinot, J. Regis, and J.-F. Mangin. Classification based on cor-
tical folding patterns. IEEE Tran. Med. Imaging, 26(4):553–
565, Jan 2007.
[4] B. Fischl, N. Rajendran, E. Busa, J. Augustinack, O. Hinds,
B. T. T. Yeo, H. Mohlberg, K. Amunts, and K. Zilles. Cor-
tical folding patterns and predicting cytoarchitecture. Cereb
Cortex, 18(8):1973–80, Aug 2008.
[5] B. Fischl, M. I. Sereno, and A. M. Dale. Cortical surface-
based analysis. II: Inflation, flattening, and a surface-based
coordinate system. NeuroImage, 9(2):195–207, Feb 1999.
[6] C. Holmes, D. MacDonald, J. Sled, A. Toga, and A. Evans.
Cortical peeling: CSF/grey/white matter boundaries visual-
ized by nesting isosurfaces. Lecture notes in computer sci-
ence, 1131:99–104, Jan 1996.
[7] A. Joshi, D. Shattuck, P. Thompson, and R. Leahy. Surface-
constrained volumetric brain registration using harmonic
mappings. IEEE Tran. Med. Imagng, 26(12):1657 – 1669,
Dec 2007.
[8] H. Le. Locating Frechet means with application to shape
spaces. Advances in Applied Probability, 33(2):324–338, Jan
2001.
[9] A. Leow, C. Yu, S. Lee, S. Huang, and H. P. . . . . Brain struc-
tural mapping using a novel hybrid implicit/explicit frame-
work based on the level method. NeuroImage, 24:910–927,
Jan 2005.
[10] G. Lohmann, D. Y. von Cramon, and A. C. F. Colchester.
Deep sulcal landmarks provide an organizing framework for
human cortical folding. Cereb Cortex, 18(6):1415–1420, Jan
2008.
[11] F. Losasso, H. Hoppe, S. Schaefer, and J. Warren. Smooth
geometry images. ACM SIGGRAPH Symposium on Geome-
try processing, 43:138–145, Jan 2003.
[12] J. Mangin, D. Riviere, A. Cachia, E. Duchesnay, Y. Coin-
tepas, D. Papadopoulos-Orfanos, P. Scifo, T. Ochiai,
F. Brunelle, and J. Regis. A framework to study the corti-
cal folding patterns. NeuroImage, 23:S129–S138, Jan 2004.
[13] X. Tao, X. Han, M. Rettmann, and J. Prince. Statistical study
on cortical sulci of human brains. Lecture notes in computer
science, pages 475–487, Jan 2001.
[14] P. Thompson and A. Toga. A surface-based technique for
warping 3-dimensional images of the brain. IEEE Tran. Med.
Imagng, 15(4):402–417, Jan 1996.
[15] M. Vaillant and C. Davatzikos. Finding parametric represen-
tations of the cortical sulci using an active contour model.
Medical Image Analysis, 1(4):295–315, Jan 1997.